Knowledge Graphs in Manufacturing and Production: A Systematic Literature Review

1 Introduction

• Knowledge graphs support downstream tasks and are already used across several domains, while manufacturing and production applications have received little attention.KGs model domain knowledge; their uses include question answering, information extraction, and recommender systems.Source: §1 · pp. 2–3
• The paper treats a knowledge graph as a node-and-edge structure and includes studies accepted in peer-reviewed venues or conforming to a reviewed KG definition.It notes that no formal definition is accepted across the entire KG community.Source: §1 · p. 2
• RDF triples and property graphs are identified as the two most popular graph data models discussed in the review.The paper also notes that most public KGs use RDF triples.Source: §1 · pp. 2–3
• The review asks what role knowledge graphs play in manufacturing and production and analyzes bibliometric facts, research types, KG characteristics, and application scenarios.It also aims to identify existing standards and consensus about KG structure and construction in the domain.Source: §1 · p. 3

2 Related Work

• The authors report no systematic literature review or mapping study dedicated to knowledge graphs in manufacturing and production.They distinguish this gap from existing surveys, reviews, and books on KG technologies generally.Source: §2 · p. 4
• Earlier surveys address KG embeddings, knowledge retrieval, enterprise KG construction, refinement, and applications such as recommendation and question answering.Examples include surveys of latent and graph-feature models for predicting graph facts and of enterprise KG architecture and construction.Source: §2 · p. 4
• Other reviews cover KG construction, literal-aware embeddings, dynamic graphs, domain-specific KGs, preprocessing, and publicly available cross-domain graphs.The cited literature also includes broad introductions to KGs and neural-network approaches to generating embeddings.Source: §2 · p. 4

3 Research Method

• The review follows the planning, conducting, and reporting phases of an SLR, whose structured method reduces—but does not eliminate—review bias.The authors note that SLR methodology does not protect against publication bias in primary studies.Source: §3 · p. 5

3.1 Research Questions

• Four research questions organize the review around bibliometrics, research maturity, KG construction techniques, and application scenarios.The application questions cover manufacturing domains, concrete use cases, and systems developed using KGs.Source: §3.1 · pp. 5–6

3.2 Data Sources and Search Strategy

• The search string combines KG-method terms with manufacturing and production synonyms, while screening requires a concrete domain application and uses NAICS as a reference.Searches covered academic databases and Google Scholar; full-text screening excluded papers without a concrete application scenario.Source: §3.2 · pp. 6–8

3.3 Data Extraction and Synthesis

• Two authors independently extracted data from each selected study, cross-checked the results, and resolved differences through discussion and consensus.Extracted items cover publication distributions, research types, KG construction, and application scenarios.Source: §3.3 · p. 8

3.4 Evaluation

• A six-point Likert scale assesses study quality, with research types weighted by evidence quality from opinion papers to evaluation research.The final numerical evaluation ranges from 0 to 1 and is intended to indicate how thoroughly studies address KG aspects.Source: §3.4 · p. 9

4.1 Results of the Search

• 24 primary studies remained after screening 833 records, with duplicate removal, relevance and language screening, author agreement, and reference snowballing shaping the selection.The abstract screen excluded 682 studies; 58 full texts were reviewed, and two studies were added by snowballing.Source: §4.1 · pp. 9, 11

4.2.1 RQ1: What are the bibliometric key facts of KG for publications?

• Studies began in 2016, rose sixfold in 2017, dipped in 2018, and more than doubled in 2019; Germany and China led by author affiliation, with 2020 counts incomplete.The review covered only the first two months of 2020, so its count was disregarded; the authors expected the upward trend to continue.Source: §4.2.1 · pp. 11–12, Figure II · p. 12, Figure III · p. 13
• Half of the studies appeared in conference proceedings and one-third in journals, while Computer Science and Engineering accounted for 19 and 9 studies, respectively.The Scimago classification is nonexclusive; the review also found three studies in Business, Management and Accounting and two in each of the remaining listed fields.Source: §4.2.1 · pp. 12, 14, Figure IV · p. 13, Figure V · p. 14

4.2.2 RQ2: Which research type facets do the identified publications address?

• 11 studies were solution proposals without evaluation or validation, compared with 9 validation studies, 2 evaluation studies, and one each for experience and opinion papers.The review therefore found that proposed solutions had not yet been widely applied in real-world settings.Source: §4.2.2 · p. 14, Figure VI · p. 15

4.2.3 RQ3: What are the specific knowledge graph characteristics?

• Top-down construction was used for 69.57% of KGs and bottom-up for 30.43%; the review found no study using machine learning or deep learning to construct a KG.The paper describes top-down schemas as restrictive for adding triples and bottom-up construction as faster with less human effort but potentially lower quality.Source: Figure VII · p. 16, §4.2.3 · p. 16
• 87.5% of KGs used RDF graphs and 12.5% used property graphs, despite literature characterizing knowledge graphs as RDF graphs.Source: §4.2.3 · p. 16, Figure VIII · p. 17

4.2.4 RQ4: What are the application scenarios of knowledge graphs?

• Half of the studies reported no concrete application domain; machinery manufacturing accounted for 4 studies, and chemical and transportation equipment manufacturing for 3 each.Fabricated metal product manufacturing and textile product mills each accounted for one study.Source: §4.2.4 · p. 17
• Reported applications included sales forecasting, production-line digital twins that adapt to environmental changes, matching production problems to knowledge, and searching for manufacturing services.Source: §4.2.4 · pp. 17–18
• [S23] uses a KG containing environment data, robot capabilities, task templates, and goal descriptions to generate action plans for autonomous robots.Source: §4.2.4 · p. 19
• [S24] uses a KG as the foundation for machine learning based on graph embeddings.Source: §4.2.4 · p. 19
• Knowledge fusion is the leading use case: 12 studies (50%) use KGs to integrate data from different information sources.Three studies each address integration of separated manufacturing processes, digital twins, and automated source-code generation; other identified uses include a KG management platform and predictive analytics, while one study had no concrete use case.Source: §4.2.4 · p. 20, Figure X · p. 20
• Search-based applications were the most common systems developed on KGs, appearing in 8 cases.Other systems included code generation (3); visualization and analysis, digital twins, cyber-physical-system modelling, simulation platforms, and automated KG construction (2 each); and KG management (1). Two studies developed no concrete KG-based system.Source: §4.2.4 · pp. 20–21, Figure XI · p. 21

5 Discussion

• The discussion examines the review results and identifies gaps and opportunities for future research.Source: §5 · p. 21
• The discussion also addresses threats to the validity of the study.Source: §5 · p. 21
• Table XII links knowledge-graph-based system kinds with use cases and studies.Source: Table XII · p. 22

5.1 About the bibliometric key facts of KG for publications (RQ1)

• Most selected studies appeared in conference proceedings, which the authors regard as an indicator of an emerging research topic.They note that journal publications with archival value often increase as fields mature.Source: §5.1 · p. 22
• Germany and China originated the most primary studies, while the United States ranked fourth and Japan had none identified.Source: §5.1 · p. 22
• Because the topic is emerging and the number of primary studies is low, the origin-country analysis is a current snapshot that may change.Source: §5.1 · p. 22

5.2 About the research type facets that the identified publications address (RQ2)

• Validation research with sound evidence remains limited, and most proposed solutions have not been implemented in real-world settings.The review notes that validation studies nevertheless provide foundations for employment of knowledge graphs in companies.Source: §5.2 · p. 23
• Experience papers and opinion papers are currently in the minority.Source: §5.2 · p. 23
• Deep-learning approaches have accelerated fully automated proposals, but this has not yet brought new views on their potential impact in manufacturing and production.The authors suggest this could change as methods for processing and reasoning about numerical and tabular data advance.Source: §5.2 · p. 23

5.3 About the specific knowledge graph characteristics (RQ3)

• Industry knowledge graphs are usually top-down and use property graphs more often than knowledge graphs in general, while the sector makes little use of embeddings.The authors suggest that a closed-world approach may fit production tasks with conditional restrictions; advanced embedding-based cognitive tasks are consequently not widely exploited.Source: §5.3 · pp. 23–24
• A few innovative works let knowledge graphs evolve freely without prespecifying structure, allowing discovery of potentially interesting facts or relationships.Source: §5.3 · p. 24
• Few papers use machine learning or deep learning to expand knowledge graphs, which the authors tentatively connect to explainability concerns and challenges with textual data.They also note that available tools may have unrealistic assumptions for industrial numerical information.Source: §5.3 · p. 24
• Most proposed solutions remain conservative and make little use of newer machine-learning developments, though the authors anticipate bottom-up graphs paired with more mature deep learning in unexplored domains.Source: §5.3 · p. 24

5.4 About the application scenarios (RQ4)

• 50% of primary studies had no explicit application domain, often because research was early-stage or proposed broadly applicable solutions.Production-line data analysis, for example, can apply across manufacturing domains.Source: §5.4 · p. 24
• Knowledge fusion is the main use case, appearing in half of analyzed studies and serving as a foundation for digital twins, process integration, program generation, and predictive analytics.The review links its prominence to the integration of different data sources.Source: §5.4 · p. 24
• Search-based applications appeared in 8 studies, while data visualization and analysis systems appeared in 2 studies, supporting exploration and knowledge discovery.Source: §5.4 · p. 25
• Three studies developed code-generation systems, including two for autonomous-robot or vehicle source code and one for generating test cases from specifications.Other systems represented manufacturing processes through digital twins, automated cyber-physical-system modeling, or simulation.Source: §5.4 · p. 25
• Three studies focused on knowledge-graph construction or management, and the review describes the overall topic as emerging with multiple open research challenges.Two investigated automated graph construction from existing data, and one proposed a graph-management platform.Source: §5.4 · p. 25

5.5.1 ORC1: Proper handling of numeric and tabular data

• Manufacturing solutions largely target text, leaving numeric and tabular machine data less addressed despite its prevalence in production environments.Integrating such data into knowledge graphs could support data analysis, integration, and knowledge discovery.Source: §5.5.1 · pp. 25–26

5.5.2 ORC2: Further research in real-time knowledge graphs

• Few studies address temporal processing, while automated manufacturing needs synchronization and fast knowledge-graph updates as information arrives from multiple sources.The authors call for further research on real-time updating.Source: §5.5.2 · p. 26

5.5.3 ORC3: Automatic linkage with other publicly available knowledge graphs

• Manufacturing knowledge graphs are often siloed rather than linked to public graphs, leaving their potential for growth underused and making automatic entity alignment a research need.The authors note that entity-alignment methods have recently appeared.Source: §5.5.3 · p. 26

5.5.4 ORC4: A higher degree of coverage of all manufacturing and production domains

• Existing approaches remain weakly validated in manufacturing, and most NAICS categories remain largely unexplored in knowledge-graph research.The authors expect evaluation research to expand across NAICS categories.Source: §5.5.4 · p. 27

5.5.5 ORC5: Collection of best practices concerning the development of reference architectures for KG development

• The field lacks collected best practices and reference architectures, so teams have developed solutions without shared guidelines.The authors argue that common practices could support quality and save development time and money.Source: §5.5.5 · p. 27

• All three authors assessed selected studies using defined criteria, then discussed and re-evaluated studies with conflicting votes.Source: §5.6 · p. 27
• A multi-stage screening process and additional data sources and search terms were used to reduce risks from study exclusion, low statistical power, and selection or instrumentation.The search used five academic libraries and Google Scholar, with alternative terms for manufacturing and production subfields.Source: §5.6 · pp. 27–28

• The review classifies primary studies by bibliometric facts, research type, knowledge-graph characteristics, and application scenarios to assess the field and identify challenges.Source: §6 · p. 28
• Almost 90% of research comes from one of the top nine manufacturing countries; most papers appear in conferences, and more than 95% underwent review.Source: §6 · p. 28
• More than 80% of selected studies are solution proposals or validation research, while evaluation papers are scarce.Source: §6 · p. 28
• 87.5% of relevant knowledge graphs are RDF graphs, while 12.5% are property graphs.Source: §6 · p. 28
• Research is fragmented across narrow sectors, leaves many NAICS categories unexplored, and rarely exploits deep-learning methods for graph processing.Source: §6 · p. 28
• The authors describe the field as still growing and expect it to remain active, while noting that it has not yet reached its peak.Source: §6 · p. 29

• Another study title concerns event-driven knowledge-graph completion in digital factories [S2].Source: Primary Studies · pp. 33–34
• A study title describes knowledge fusion of manufacturing operations data using representation learning [S3].Source: Primary Studies · p. 33
• Listed work also includes knowledge-graph-based digital twin models for industrial production lines [S4].Source: Primary Studies · p. 33